Hyperparameter selection of one-class support vector machine by self-adaptive data shifting

With flexible data description ability, one-class Support Vector Machine (OCSVM) is one of the most popular and widely-used methods for one-class classification (OCC). Nevertheless, the performance of OCSVM strongly relies on its hyperparameter selection, which is still a challenging open problem due to the absence of outlier data. This paper proposes a fully automatic OCSVM hyperparameter selection method, which requires no tuning of additional hyperparameter, based on a novel self-adaptive “data shifting” mechanism: Firstly, by efficient edge pattern detection (EPD) and “negatively” shifting edge patterns along the negative direction of estimated data density gradient, a constrained number of high-quality pseudo outliers are self-adaptively generated at more desirable locations, which readily avoids two major difficulties in previous outlier generation methods. Secondly, to avoid time-consuming cross-validation and enhance robustness to noise in the given training data, a pseudo target set is generated for model validation by “positively” shifting each given target datum along the positive direction of data density gradient. Experiments on synthetic and benchmark datasets demonstrate the effectiveness of the proposed method.